Analysis of continuous dynamical system models with Hessians derived from optimization methods

From MaRDI portal





The authors consider continuous-time dynamical systems involving Hessians, derived from unconstrained continuous optimization methods, \(x_{\ast}\in\arg\min_{x\in \mathbb{R}^{n}}f(x)\), of the form: \[ a_{1}\ddot{x}+a_{2}\dot{x}+a_{3}\nabla^{2}f\dot{x}+a_{4}\nabla f=0 \tag{1} \] where \(a_{i}:\mathbb{R}^{+}\rightarrow \mathbb{R}\) \((i=1,2,3,4)\) are some functions of \(t\), and \(a_{3}\) and \(a_{4}\) are not zeros. They classify \((1)\) into three cases, and for each case, they exhibit a Lyapunov function and obtain the best rate of convergence when using this Lyapunov function. However, in some cases, this rate is the same as either the simplest gradient flow, or the continuous dynamical system corresponding to Nesterov's accelerated gradient method. Therefore in these cases, the incorporation of the Hessian does not bring any improvement. Their results extend previous results in [\textit{H. Attouch} et al., J. Differ. Equations 261, No. 10, 5734--5783 (2016; Zbl 1375.49028); \textit{B. Shi} et al., Math. Program. 195, No. 1--2 (A), 79--148 (2022; Zbl 1500.65026); \textit{J. J. Suh} et al., ``Continuous-time analysis of accelerated gradient methods via conservation laws in dilated coordinate systems, Preprint, \url{arXiv:2202.05501}].











This page was built for publication: Analysis of continuous dynamical system models with Hessians derived from optimization methods

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q6601496)